The processors in general-purpose computers, as well as those used as embedded controllers, are typically programmed to handle a plurality of tasks concurrently. A subset of these tasks must be performed in a timely manner, in response to specific, exogenous events, while the remainder of these tasks can be performed without stringent, real-time constraints. To handle both sets of tasks using a single data path, these processors require an efficient mechanism for responding rapidly to exogenous events, while allowing non-real time processing to occur whenever no exogenous events are being handled.
The predominant mechanism for event response is program interruption, which was first used in the mid-1950s. For the past 40 years, the vast majority of processor architectures have included a program interruption facility that suspends the execution of a "background" task, and initiates the execution of a "foreground" task, upon occurrence of the exogenous event(s). Each program interruption, typically called an "interrupt," causes a reversible change to the execution state of the processor upon assertion (suitably synchronized to the processor's instruction flow) of an appropriate event.
The priority interrupt, developed in the late-1950s, is a common enhancement to a program interruption facility. In a processor supporting priority interrupts, discrete priorities are assigned, either statically or dynamically, to a plurality of event (interrupt request) signals. Associated with each of these signals is a uniquely identifiable resultant state for the reversible change in execution state of the processor. Each occurrence of a priority interrupt selects the resultant state associated with the highest priority interrupt request asserted at the time when the interrupt state change is initiated.
The fundamental action when performing a reversible change in the program execution state of a processor is to save the interrupted program's execution address (and implicit inter-instruction status, such as condition codes), and to commence interrupt processing at a program address associated with the event causing the interruption. This program address is generally obtained from a predetermined memory location known as an interrupt vector. At the end of the interrupt handling routine, the saved execution address (and status value, if any) are restored, permitting execution of the interrupted program to resume at the point of interruption. In most interrupt handling routines, it is necessary to save, and subsequently to restore, additional processor state to perform the operations necessary to respond to the interrupt. This additional state is primarily the contents of processor registers other than the program counter.
Saving and restoring these registers to/from a stack or dedicated block of memory can consume considerable amounts of time. Therefore, as integrated circuits began reducing the cost and size of hardware registers in the mid-1960s, some processors were equipped with multiple sets of registers. Selection of a different set of registers, either by the interrupt support hardware or by the interrupt handling software, allowed substantially faster interrupt response by eliminating the overhead of saving and restoring registers to/from main memory.
The multiple register set concept reached its modern form on the IBM System/7, introduced in 1970. The System/7 had a dedicated, hardware-selected register set for each interrupt level, and reduced interrupt context switching time still further by including in each set a register to save the execution address (program counter value) when the level was preempted by an interrupt on a higher priority level. The result was an interrupt context switch time of 800 ns and an interrupt return time of 400 ns, both of which were truly exceptional speeds for a 16-bit minicomputer built using 1969 technology. The System/7 also pioneered dynamic interrupt assignment, where the priority level used by each interrupt source was set by software, and could be changed during system operation.
The ultimate generalization of this register set plus program counter technique was to allow events to initiate handling routines at their last execution address, rather than requiring them always to start using an interrupt vector address. For controlling I/O devices, data communication and network protocols, and other processes defined in terms of communicating state machines, this was a major benefit, because a state machine could be implemented using the level's program counter both for instruction addressing and as the (implicit) state register. This not only eliminated the need for a separate state register, but also eliminated the overhead of a dispatch routine to select the appropriate handling routine based on the value in the state register. In effect, the register set plus program counter architecture provides direct hardware support for the "task" or "execution thread" concepts commonly supported by operating system software.
The first machine developed with the intent to implement I/O control state machines using this technique was the "Alto" experimental personal computer, designed in 1972 by Charles Thacker at the Xerox Palo Alto Research Center. Since the early-1970's many variations of these interrupt and context switching mechanisms have been developed for single-chip microcomputers and microprocessors. However, none of these variations have introduced a fundamentally new mechanism for rapid context switching in response to exogenous events.
In high-performance systems it is often possible to dedicate (one or more) processors for I/O control and/or external event handling. However, if implemented with similar technology to that used in the central processor(s) of the system, the utilization of these I/O processors tends to be very low. This is due to the fact that, for any particular circuit technology, the logic devices used to implement processor data paths operate significantly faster than the storage devices used to implement main memory, and both the logic and memory devices can support higher data bandwidths than any of the attached peripheral devices.
During the 1960s, the architects of high-performance systems that required multiple I/O controllers developed a technique to share a single data path among a plurality of controller functions, even though those functions are logically disjoint. The technique used a single physical data path and instruction decoder to process, on a round-robin basis, the instruction streams of a plurality of logical processors. The only dedicated resource for each logical processor was the storage to hold its execution state (program counter and register values). The control circuitry allowed execution of a predetermined number of instructions (generally 1) for each logical processor on a sequential, cyclic basis. This control circuitry changed which one of the stored execution states was accessible to the data path between the instruction cycles for different logical processors. This technique was first used by Seymour Cray in the early 1960s to implement 10 I/O controllers (called peripheral processors or "PPUs") using a single, shared data path on the Control Data Corporation (CDC) model 6600.
Note that this logical processor state switching occurred on a strict time basis, and not in response to external events. Indeed, some successors to the Control Data 6600 PPUs implemented a priority interrupt scheme on their logical processors. More recently this data path sharing technique has been applied to central processors, where it is called "shared resource multiprocessing." In this case a plurality of independent instruction streams, from different CPU tasks or programs, are interleaved to decrease pipeline dependencies, thereby improving resource utilization, of a superscalar data path.
Accordingly, what is needed in the art is a way to configure, allocate and manage contexts that has a more general flexibility.